Author: Hirschmüller
Year: 2008
- complexity linear to number of pixels and disparity range
- MI used to handle radiometric differences
- pb with occlusion: you can do left / right association
- SOA is based on global energy cost
- derive a pixel wise MI based cost function, and a sum over all pixels gives a global cost function
- This cost function needs a prior on disparity images, that is obtained recursively on ds images
- add penalty terms for local changes of disparity in the vincinity of the pixel (smoothness) that is ponderated by the intensity gradients
- such a 2D global minimization is NP complete
- aggregate the cost of all the 1D pixels costs that ends up in the pixels (to smooth again)
- subpixel refinement by interpolating a cuadratic curves on the 3 lowest costs and finding the min
- Compute the two disparity images switching base and match image and remove outliers
- performs multibaseline stereo with : a weighted mean of all pixels + remove outliers that are within 1 pixel from the median
- post processing: remove peaks, fit planes on untextured areas, interpolates occluded areas with the disparity of the occludee, interpolates missing values with a median disparity
TOCHECK stereo:
- est-ce que l'aller-retour est activé?
- combien de directions pour l'agregation de couts
- utiliser la valeurs médiane pour la fusion de costmaps